typo

dsmat.c

@@ -13,6 +13,7 @@ Matrix* alloc_matrix(int mb, int nb, int b, int p, int q, starpu_mpi_tag_t* tag)
 	X = malloc(sizeof(Matrix));
       	X->blocks = malloc( mb*nb*sizeof(Block));
 	int i,j;
+//	printf("[%d] allocating %d x %d B %d on %dx%d\n", comm_rank, mb, nb, b, p, q);
 	for (i = 0; i<mb; i++)
 	{
 		for (j= 0; j<nb; j++)
@@ -54,7 +55,7 @@ void free_matrix(Matrix* X)
 	free(X);
 }
 
-void register_matrix(Matrix* X, int mb, int nb, int datatype, int prune_handles, int p, int q, int row, int col, int check, int delay)
+void register_matrix(Matrix* X, int mb, int nb, char* name, int datatype, int prune_handles, int p, int q, int row, int col, int check, int delay)
 {
 	starpu_mpi_comm_rank(MPI_COMM_WORLD, &comm_rank);
 	int proc_row, proc_col;
@@ -64,14 +65,15 @@ void register_matrix(Matrix* X, int mb, int nb, int datatype, int prune_handles,
 	// comm_rank = proc_row * q + proc_col
     	proc_col = comm_rank % q;
 	proc_row = (comm_rank - proc_col)/q; 
-//	printf("[%d] delayed ? %d / check ? %d / dt ? %d / pr_hd ? %d\n", comm_rank, delay, check, datatype, prune_handles);
+//	printf("[%d = %d x %d | %d x %d] delayed ? %d / check ? %d / dt ? %d / pr_hd ? %d\n", 
+//		comm_rank, proc_row, proc_col, row, col, delay, check, datatype, prune_handles);
 	for (b_row = 0; b_row < mb; b_row++)
 	{
 		for (b_col = 0; b_col < nb; b_col++)
 		{
 			Xij = & X->blocks[b_row*nb+b_col];
 			Xij->hdl = malloc(sizeof(starpu_data_handle_t));
-//			printf("[%d] X_%d,%d | tag:%d\n",comm_rank,b_row,b_col,Xij->tag);
+//			printf("[%d] %s_%d,%d=%dx%d | tag:%d | owned by %d\n",comm_rank,name,b_row,b_col,b_row%p,b_col%q,Xij->tag,Xij->owner);
 	    		if (Xij->owner == comm_rank)
 			{
 				if (datatype) {
@@ -90,7 +92,7 @@ void register_matrix(Matrix* X, int mb, int nb, int datatype, int prune_handles,
 						     (check && comm_rank == 0)) ) {
 				block_starpu_register(Xij, datatype);
 			} else {
-//				printf("[%d] pruned X_%d,%d\n",comm_rank,b_row,b_col);
+//				printf("[%d] pruned %s_%d,%d\n",comm_rank,name,b_row,b_col);
 			}
 		}
 	}

dsmat.h

@@ -19,7 +19,7 @@ typedef struct Matrices
 
 Matrix* alloc_matrix(int mb, int nb, int b, int p, int q, starpu_mpi_tag_t * tag);
 void free_matrix(Matrix* X);
-void register_matrix(Matrix* X, int mb, int nb, int datatype, int prune_handles, int p, int q, int row, int col,int check, int delay);
+void register_matrix(Matrix* X, int mb, int nb, char* name, int datatype, int prune_handles, int p, int q, int row, int col, int check, int delay);
 void unregister_matrix(Matrix* X, int mb, int nb);
 void print_matrix(Matrix* X, char* name);
 void block_starpu_register(Block* Xij, int datatype);

starpu_example_dgemm.c

@@ -214,15 +214,15 @@ static void register_matrices()
 		starpu_tile_interface_register();
 	}
 
-	register_matrix(A,MB,KB,datatype,prune_handles,P,Q,1,0,check,delay);
-	register_matrix(B,KB,NB,datatype,prune_handles,P,Q,0,1,check,delay);
+	register_matrix(A,MB,KB,"A",datatype,prune_handles,P,Q,1,0,check,delay);
+	register_matrix(B,KB,NB,"B",datatype,prune_handles,P,Q,0,1,check,delay);
 	//register_matrix(C,C_h,&tag,MB,NB,datatype,prune_handles,P,Q,0,0); 
 	// FIXME (starpu-side) :
 	// the previous one seems logical because we do not need to know 
 	// about blocks of C we do not contribute to, however startPU seems to be
 	// pending on task_insertion if we do not know about blocks on our row/column even if we do not contribute to them)
 	// - This could happen because we are seeing a Write on a NULL handle and StarPU is waiting (for what ?)
-	register_matrix(C,MB,NB,datatype,prune_handles,P,Q,1,1,check,delay);
+	register_matrix(C,MB,NB,"C",datatype,prune_handles,P,Q,1,1,check,delay);
 }
 
 /* Unregister matrices from the StarPU management. */
@@ -604,7 +604,7 @@ int main(int argc, char *argv[])
 		Matrix* Cwork;
 		if (check) {
 			Cwork = alloc_matrix(MB,NB,BS,1,1,&tag);
-			register_matrix(Cwork,MB,NB,datatype,0,1,1,1,1,1,0);
+			register_matrix(Cwork,MB,NB,"Cwork",datatype,0,1,1,1,1,1,0);
 			copy_matrix(Cwork,C);
 			starpu_mpi_wait_for_all(MPI_COMM_WORLD);
 			if (verbose) print_matrix(C,"Cinit");
@@ -625,7 +625,7 @@ int main(int argc, char *argv[])
 				{
 					Ail = & A->blocks[b_row*KB + b_aisle];
 					Blj = & B->blocks[b_aisle*NB + b_col];
-					Cij = & C->blocks[b_col * NB + b_row];
+					Cij = & C->blocks[b_row * NB + b_col];
 				 	a_local = Ail->owner == comm_rank;	
 				 	b_local = Blj->owner == comm_rank;	
 				 	c_local = Cij->owner == comm_rank;	
@@ -648,8 +648,14 @@ int main(int argc, char *argv[])
 							clargs = malloc(sizeof( struct cl_zgemm_args_s ));
 							clargs->alpha = alpha;
 							clargs->beta = b_aisle==0? beta : 1.0;
+						} else {
+//							printf("[%d] inserting %d,%d (l:%d)\n",comm_rank,b_row,b_col,b_aisle);
 						}
-						starpu_mpi_task_insert(MPI_COMM_WORLD, &gemm_cl,
+/*						printf("[%d] C_%d,%d (%d-%p-%d) += A_%d,%d (%d-%p-%d) B_%d,%d (%d-%p-%d)\n", comm_rank,
+							b_row,b_col,  Cij->registered,Cij->hdl,Cij->owner,
+							b_row,b_aisle,Ail->registered,Ail->hdl,Ail->owner,
+							b_aisle,b_col,Blj->registered,Blj->hdl,Blj->owner);
+*/						starpu_mpi_task_insert(MPI_COMM_WORLD, &gemm_cl,
 							STARPU_CL_ARGS, clargs, sizeof(struct cl_zgemm_args_s), 
 							STARPU_R, *Ail->hdl,
 							STARPU_R, *Blj->hdl,
@@ -681,9 +687,9 @@ int main(int argc, char *argv[])
 			Bcheck = alloc_matrix(KB,NB,BS,1,1,&tag);
 			Ccheck = alloc_matrix(MB,NB,BS,1,1,&tag);
 
-			register_matrix(Acheck,MB,KB,datatype,0,1,1,1,1,1,0);
-			register_matrix(Bcheck,KB,NB,datatype,0,1,1,1,1,1,0);
-			register_matrix(Ccheck,MB,NB,datatype,0,1,1,1,1,1,0);
+			register_matrix(Acheck,MB,KB,"Acheck",datatype,0,1,1,1,1,1,0);
+			register_matrix(Bcheck,KB,NB,"Bcheck",datatype,0,1,1,1,1,1,0);
+			register_matrix(Ccheck,MB,NB,"Check",datatype,0,1,1,1,1,1,0);
 
 			copy_matrix(Acheck,A);
 			if (verbose) print_matrix(A,"A");